06:15 <fijal> mattip: mail richard? or whoever is handling that

09:34 <fijal> https://www.irccloud.com/pastebin/LoedRcgo/

09:34 <fijal> server

09:34 <fijal> https://www.irccloud.com/pastebin/i1qihBFq/

09:34 <fijal> client

09:35 <fijal> cfbolz: ok, so the "allocating the buffer not in a loop" is the biggest thing here

09:39 <fijal> it's probably worth asking though why this is 40x faster than a urllib client request though

10:32 <fijal> ok, I don't get it, ssl just broke on baroquesoftware

10:51 <fijal> https://www.irccloud.com/pastebin/PeAzlswb/

10:59 <cfbolz> fijal: the difference in speed comes from line 39?

11:00 <fijal> yes

11:01 <cfbolz> fijal: that's a great benchmark, I'd say (didn't run it yet, but will in a bit)

11:02 <fijal> its a bit of a mess to set up

11:02 <fijal> and maybe exception catching is bad (but also fuck, why is this so hard)

12:05 <cfbolz> fijal: what's a bit annoying is that without warmup cpython is significantly faster

12:09 <cfbolz> on my laptop commenting in line 39 is even a ~600x slowdown for cpython???!

12:13 <LarstiQ> handwaving something about `sendfile(2)`?

12:14 <cfbolz> no, something really weird is happening

12:17 <cfbolz> I mean, this pattern does really not look good:

12:17 <cfbolz> https://www.irccloud.com/pastebin/dE7afnSj/

12:18 <cfbolz> (sends requests in bunches of 100, sometimes it takes 0.001s, sometimes 3s)

12:26 <cfbolz> seems to correlate with the tls connection dropping

12:52 <fijal> i run more per loop

12:52 <fijal> but you need a localhost web server i think

13:09 <cfbolz> Right

13:09 <cfbolz> Later

13:10 <cfbolz> fijal: anyway, if the connection drops, the numbers are meaningless

13:10 <fijal> yeah

14:23 <ctismer_> cfbolz: Hi! Was there something changed between PyPy3.7 and 3.8? I have problems building PySide

14:24 <ctismer_> cfbolz: (BTW.: although there are errors, there are so many examples running that we announce it, now :) )

14:39 <ctismer_> cfbolz: no, was just a setuptools glitch. Distutils was more `idempotent`.

14:40 <cfbolz> ctismer_: cool!

14:40 <cfbolz> Where are you announcing it btw?

14:40 <cfbolz> Do you think it's far enough that we should write a blog post about it?

14:43 <ctismer_> cfbolz: ATM only in our documentation pages, as new feature for PySide 6.3 (comes in a few weeks). I will tell the story and show what works and what not, how to install and so on. I also still have trouble installing scipy stuff.

14:43 <cfbolz> ctismer_: excellent, let's wait for the release then

14:45 <ctismer_> cfbolz: Yes, I think it is far enough. Did not realize that when desperately searching why certain signals do not work, but we got only 9 of 134 examples not working.

14:47 <ctismer_> cfbolz: One funny finding was `QPainter`. That object is often created without closing it, creating many errors in PyPy. Until I recognized that this should be a contest manager since many years 😂

14:48 <cfbolz> ctismer_: that sounds like progress

14:49 <ctismer_> cfbolz: context manager, that automatically calls painter.end()

14:50 <ctismer_> Yes. I'm waiting for the next version, so I can get certain extensions without building.

14:51 <ctismer_> (because scipy and sci-image drive me crazy and conda collides too (trying since days))

14:57 <cfbolz> That's beyond my pay grade 😂

15:26 <mattip> the problem with getting pypy into scipy CI is that tests run very slow

15:26 <mattip> most of the code uses cython/pythran/pybind11

15:26 <mattip> and it is hard to convince them to release a wheel if they don't run the tests

15:26 <mattip> (justifyably)

16:08 <ctismer_> mattip: And with wheels it would work fine, but I need the nightly build until my changes are officially in.

16:10 <ctismer_> BTW., is it planned to have PyPy 3.9 in the release, or will it be 3.8, still?

16:17 <cfbolz> ctismer_: 3.8, plus 3.9 alpha

16:17 <ctismer_> cfbolz: ok, so I should stick with 3.8 for now.

16:18 <cfbolz> ctismer_: yep